Research · Preprint, 2026
Reading the Architecture off the Weights: Manifest-Only Recovery of Neural Network Anatomy
Paper on ResearchGate Try AI Model Explorer Code (MIT) Plain-language explainer
Abstract
Engineers who deploy a model need to know what it is made of: which blocks repeat, what each block computes, how the parts connect and where the parameters sit. This information is usually spread over model cards, family-specific configuration files and modeling code. We show that most of it can be read from the tensor manifest alone, the list of tensor names, shapes and storage types that a safetensors file keeps in a header of a few tens of kilobytes. Our method normalizes quantized storage, finds repeated blocks in the module tree, assigns roles to the parts of each block from the algebra of their shapes, and links components by matching their widths. Fields that shapes cannot determine, such as the number of attention heads, are read from config.json when it exists, and the configuration is checked against the weights. On 510 architectures from the transformers library, with ground truth taken from module classes that the method never sees, per-tensor role accuracy is 96.8%, against 82.3% for the strongest name-rule baseline, and 95.0% when every name is replaced by a random token. Parameter counts are exact for every model, including under NVFP4, FP8 and GPTQ storage. On 191 popular Hugging Face checkpoints read over HTTP range requests, totals agree with the Hub's own count wherever the Hub counts logical parameters, and configuration fusion recovers head geometry for 96 to 98% of models. The method runs in about 10 ms for a typical model, and we release it as an open explorer that reads any Hugging Face repository without downloading weights.
Who created this
Sattaya Singkul created the manifest-only recovery method and AI Model Explorer. The tool began with the workshop “Squeeze Until It Breaks: A Hands-On Edge Inference Contest” at AI Expo Thailand 2026, was then generalised to any architecture, and is evaluated in this paper. The checkpoint benchmark is public.
Cite
@misc{singkul2026reading,
title = {Reading the Architecture off the Weights: Manifest-Only Recovery of Neural Network Anatomy},
author = {Singkul, Sattaya},
year = {2026},
note = {Preprint},
howpublished = {ResearchGate},
url = {https://www.researchgate.net/publication/417293353_Reading_the_Architecture_off_the_Weights_Manifest-Only_Recovery_of_Neural_Network_Anatomy}
}